Claude Fable 5 is back with tighter cybersecurity blocks

Claude Fable 5 is back with tighter cybersecurity blocks.

Anthropic revealed more details on Fable 5's cyber safeguards and its jailbreak framework.

Just in case you missed the story, here's what happened. Amazon found a bypass to Fable 5's safety filters making the model find vulnerabilities and write an exploit.

The US government treated that as a national security issue and forced Anthropic to revoke access to the model using export control measures.

Anthropic strengthened its safeguards to return the model back:

  • A new safety classifier with lower blocking and rerouting thresholds.
  • Some routine coding and debugging requests will temporarily fall back to Opus 4.8 while Anthropic tunes the filters to cut false positives.
  • A new cross-industry jailbreak severity framework is being drafted with Amazon, Microsoft, Google, and other Glasswing partners.
Fable 5 widens the classifier safety margin, blocking more benign requests. Source: Anthropic.
Fable 5 widens the classifier safety margin, blocking more benign requests. Source: Anthropic.

My take:

  1. Anthropic's marketing move backfired. On one hand, Anthropic markets Fable as a dangerously capable model, but at the same time, they had to say that "every model we tested could produce the same demonstration as Fable 5".
  2. The deployed classifier and thresholds will come at the cost of very high false positives. I just tried to summarize a news article and got "Fable 5's safeguards flagged this message. The safeguards are intentionally broad right now and may flag safe and routine coding, cybersecurity, or biology work."
  3. Frontier labs are hit with the real problem of ensuring that the good guys can use the dual-use capabilities, but the bad guys can't. And this problem is not really solvable with classifiers.

Sources:

More details on Fable 5's cyber safeguards and our jailbreak framework, Anthropic